Repository navigation
Seven cast rules push nullability casts without trivially_cast_nullability #10750
Description
Problem or motivation
The cast-rule contract on Validity::cast_nullability (vortex-array/src/validity.rs) asks a CastReduce rule to call Validity::trivially_cast_nullability, to return Ok(None) when that call cannot prove a nullability cast safe, and to pair the rule with a CastKernel for execution. These cast rules instead push a nullability cast into the child that holds their validity, with no guard:
| Encoding | Rule | Child |
|---|---|---|
| RunEnd | encodings/runend/src/compute/cast.rs |
values |
| ZigZag | encodings/zigzag/src/compute/cast.rs |
encoded |
| FoR | encodings/fastlanes/src/for_/compute/cast.rs |
encoded |
| RLE | encodings/fastlanes/src/rle/compute/cast.rs |
indices |
| ALP | encodings/alp/src/alp/compute/cast.rs |
encoded |
| ALP-RD | encodings/alp/src/alp_rd/compute/cast.rs |
left parts |
| DateTimeParts | encodings/datetime-parts/src/compute/cast.rs |
days |
From reading the code, no wrong result is visible today, because the null check still runs when the child cast executes. The rules depend on every child cast keeping that check, though. DecimalByteParts had the same gap, and #10746 fixes it with the guard plus a CastKernel.
Proposed solution
Apply the DecimalByteParts fix to each rule: guard it with trivially_cast_nullability, add a CastKernel for the case that needs execution, and test nullable to non-nullable casts with nulls, without nulls and with only nulls. A shared helper for "push a nullability cast into the validity child" can avoid seven copies of the guard.
Open question: does the contract apply to rules that do not inspect validity themselves? The contract asks for a kernel only from rules that do.
Activity
connortsui20 commentedon Oct 26, 2026
Closing as not planned. None of these rules returns a wrong result: each one pushes the nullability cast into the child that holds its validity, and the cast of that child fails when a row is null. The same pattern also shows up in Chunked, Sparse and Extension, so a guard plus a
CastKernelper encoding would add a lot of code and delay each rewrite until execution without changing any result.That answers the open question above: the guard and the kernel are for rules that build a new
Validity. #10798 documents both forms onCastReduce, withDictas the exception. Every push-down encoding already runstest_cast_conformanceon a partially valid array, which asserts the error. #10746 has the discussion for DecimalByteParts.Generated by Claude Code
Document the nullability contract of cast rules (#10798)